The hard part of a robot was never the body. It was the software that turns seeing into doing, and a lab inside one of China's biggest companies has now given that away for nothing. It runs on a gaming graphics card, and it drives 20 different kinds of robot.
Watch a humanoid robot fold a towel, and you will be watching the wrong thing.
The impressive part is not the hands. Hands are mechanical engineering, and mechanical engineering, given enough money, gets solved. Boston Dynamics solved the walking. Plenty of firms have solved the gripping.
The part that has kept robots perpetually five years away, for about thirty years, is the thing you cannot see: the software that looks at a rumpled towel, works out what a towel is, decides what folding means, and then drives forty joints in the right order to do it. That is the brain problem, and it has been brutally, expensively unsolved.
On 8 July, a lab inside China's Ant Group gave a version of it away for free.
What actually landed
Robbyant, an embodied-AI unit inside Ant Group (the company behind Alipay), released LingBot-VLA 2.0 and open-sourced it, weights and all, on GitHub and Hugging Face.
The claims, which we will hold at arm's length in a moment, are these. It is a 6-billion-parameter model trained on roughly 60,000 hours of real footage: about 50,000 hours of robots doing things, plus 10,000 hours of first-person video of humans doing things. It was trained across 20 different robot body types from 17 manufacturers. And one model can drive all of them, coordinating hands, waist, head and a wheeled base together, with under 130 milliseconds of lag, running on a single gaming graphics card.
An important correction to the version of this story you may have seen: this is not the first time LingBot has been open-sourced. Version 1 went out in late January. What changed in July is that version 2 got good enough to matter, and broad enough to matter across bodies. The news is not that China open-sourced a robot model. It is that the free one is now a serious option.
What a "vision-language-action" model actually is
Strip the acronym away and it is simple. A VLA takes two inputs and produces one output:
Vision. What the robot's cameras can see right now. A cluttered bench, a half-open drawer, a towel.
Language. A plain instruction. "Put the mug in the sink." Not code. English.
Action. The actual motor commands. Which joint moves, how far, in what order, updated many times a second.
That is the whole trick, and it is the thing that used to require a specialist team, a specialist dataset and a specialist robot. Every company built its own, from scratch, for its own machine. It was the software equivalent of every car maker inventing its own petrol.
Why "one model, twenty bodies" is the sentence that matters
Robot software has historically been agonisingly specific. A model trained on a two-armed picker was useless on a walking humanoid, because the number of joints, their arrangement and their physics were all different. Change the body, retrain everything.
A model that transfers across 20 body types breaks that. It means the software layer stops being bespoke and starts being a component, the way an operating system is a component. You build the robot; you download the brain.
And if that sounds familiar, it should. This is the Android pattern. Google did not win phones by making the best handset. It won by giving away the layer everybody needed, so that every handset maker in the world built on it, and Google set the defaults. Free is not charity. Free is a land grab.
Which raises the question sitting underneath this entire story: whose defaults?
The race nobody announced
One day after Ant's release, on 9 July, NVIDIA and Hugging Face moved. They pushed their own open robot model, Isaac GR00T 1.7, into LeRobot, Hugging Face's open robotics library, along with a teleoperation framework and an open dataset of more than 350,000 robot trajectories and 57 million grasp examples.
Two camps, one week, the same strategy. A Chinese free universal model on one side; a Western open stack of model, data and tooling on the other. Neither is charging. Both are trying to become the thing everyone builds on, because the layer everyone builds on is the layer that collects the rent later, in chips, in cloud, in ecosystem gravity.
The interesting part is that a small robotics startup in Ohio or Osaka now genuinely benefits. The piece that used to take years and a large lab is sitting on a download page. That is a real, immediate democratisation.
The uncomfortable part sits right next to it. When the default plumbing of a strategic industry is written elsewhere, that is not nothing, even when it is free. Especially when it is free.
If the model is free, where is the moat?
This is the question every robotics investor is now asking, and it has a fairly clear answer.
Data. Ant trained on 60,000 hours of footage. NVIDIA and Hugging Face released 350,000 robot trajectories. Nobody is giving away the data they gather next, from robots doing real work in real buildings. The model is the commodity; the experience is not.
Deployment. Getting a robot to fold a towel in a demo is a research problem. Getting it to do so for eight hours, beside a person, without hurting anyone or breaking anything, is an industrial one. That is where the years go.
The body. Free software makes hardware more valuable, not less, because the differentiator moves to whoever builds the most reliable, cheapest, most repairable machine.
Which is the quiet irony of this week. By giving away the brain, both camps have made the boring parts, the joints, the sensors, the safety cases, the service contracts, into the thing that actually decides who wins.
The honest catch
The figures are the company's own. The 20 body types, the 60,000 hours, the 130-millisecond latency: all vendor-stated. Independent benchmarks are not in.
But the weights are open. Which is the saving grace, and unusual. Anyone can download it and check. Claims that can be tested tend to get tested.
This is version two, not version one. Anyone telling you China just open-sourced a robot brain is a few months behind.
A model is not a robot. It still needs a body, sensors, safety engineering and someone willing to let it near a real workspace.
EDITOR'S TAKE
Ignore the robot videos. The thing to watch this year is which brain the robots are running, and the answer is drifting toward whichever one is free. Two of the most powerful technology ecosystems on earth spent the same week giving away the hardest, most expensive component in robotics, and neither did it out of generosity. They did it because the layer everyone builds on is the layer that sets the rules, and both know it. For a small robot builder this is genuinely wonderful news: the moat just got drained. For everyone else, the question is not whether robots are coming, it is whose defaults they will arrive with. That decision is being made right now, quietly, on a download page.
Quick questions
What is a vision-language-action (VLA) model?
A VLA is the software that acts as a robot's brain. It takes what the robot's cameras see, plus a plain-English instruction such as "put the mug in the sink," and turns them into the actual motor commands that move the robot's joints. It replaces the traditional approach of hand-coding behaviours for one specific machine, and it is the piece that has been hardest and most expensive to build.
Did China just open-source the first robot brain?
No, and this is a common misreading. Ant Group's LingBot-VLA was first open-sourced in late January 2026. What was released on 8 July is version 2.0, an upgrade that its makers say drives 20 different robot body types on a single gaming graphics card. The significance is not that a free robot model exists, it is that the free one has become a credible alternative to the paid ones.
Why would a company give away its robot software for free?
For the same reason Google gave away Android. The layer that every builder depends on is the layer that sets the defaults, shapes the ecosystem and captures value later through chips, cloud services and gravity. Ant Group released LingBot-VLA 2.0 free; a day later NVIDIA and Hugging Face pushed their own open robot model and dataset into the LeRobot library. Neither is charity. Both are attempts to become the standard everyone else builds on.
Sources
RoboticsTomorrow: Robbyant upgrades and open-sources LingBot-VLA 2.0.
NVIDIA: Isaac GR00T 1.7 and new frameworks come to Hugging Face's LeRobot.
MarkTechPost: LingBot-VLA 2.0, an open 6B vision-language-action model.
Related from Frontier Signal: this week's deep dive on the chip that writes DNA with electricity and water. Frontier Signal explains frontier technology in plain English.
